Prism Audio Video Bridge

What it is

The dual-tower bridge - the component that lets Prism's video and audio towers influence each other while they denoise. Not a model you prompt, a piece of wiring with weights in it: it couples the video experts and the audio expert so that a lip movement and a phoneme end up agreeing with each other instead of being two independent generations stapled together in the same file.

You reach for it because Prism Native Pipeline requires the dual_tower_bridge socket, and because this is the only node that produces it.

The boring-looking file that carries two important numbers

Here's the part nobody would guess from a dropdown with one item in it. When the sampler assembles the pipeline, it pulls two things out of this component's metadata:

• prism.boundary_ratio - the value that decides where the high-noise video expert hands over to the low-noise one. That's why the handoff between Prism High Noise Video DiT and Prism Low Noise Video DiT isn't a widget anywhere in the pack: it's baked into this file.
• prism.scheduler_config - the config the paired flow-match scheduler is built from. The pair scheduler is what keeps the video and audio schedules in step with each other, and it's constructed from what this resource declares.

So the bridge is a genuine component rather than a routing node, and swapping in a bridge from a different conversion bundle changes sampling behaviour as well as wiring. Which is a good reason to keep all seven files from one bundle.

The rest is mechanics: the loader inspects the safetensors header, confirms the declared component kind is dual_tower_bridge, and hands a descriptor to the pipeline node. Weights load at sampling time, not at graph-build time.

Input and output

model_name - the enum over your models/diffusion_models/ folders. The file is prism_alpha_dual_tower_bridge_int8_convrot.safetensors, from t8star/Prism-Comfy.

One output: dual_tower_bridge, into the socket of the same name on Prism Native Pipeline.

The quantisation story is the same as the other ConvRot files: block Linear layers are INT8, embeddings, norms and time projections stay in float. At sampling time, the bridge keeps floating-point activations even when you pick the kitchen backend - the int8_backend switch only governs the video towers.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Prism-T8.git
cd Comfyui-Prism-T8
python -m pip install -r requirements.txt

ComfyUI Manager → Comfyui-Prism-T8, or registry t8star/prism-t8, if you'd rather not babysit a terminal. Restart ComfyUI, then place the file in models/diffusion_models/ (or in the pack's own models/diffusion_models/, which is scanned too). No Triton needed for the bridge - that requirement belongs to the native sparse-attention path.

Where people get burned

Bundle soup. Components are from different Prism variants/conversion bundles. The bridge is the component whose metadata the runtime reads for the expert handoff and the scheduler, so a mismatched bridge isn't a cosmetic problem. Re-download the full seven-file set.

Empty dropdown. Either the file isn't in a scanned model folder, or it isn't a converted Prism component - the loader will tell you not a Prism standalone component. Convert official weights first. Raw MOVA-360p or Diffusers weights won't do.

Hunting for a "bridge strength" setting. There isn't one in this pack. Cross-modal influence is what the component is; you steer the audio side with audio_prompt, the audio tags (<speech>, <music>, <sfx>) and the sampler's audio_shift, not by tweaking the bridge.

Blaming the bridge for bad sync. If speech and lips disagree, the likely culprit is downstream, not here - the author's own reproduction found input-dependent audio degradation in Prism's base model itself. Check the soundtrack with your ears before you start re-downloading weights.